1 // Copyright 2014 Google Inc. All Rights Reserved. 2 // 3 // Use of this source code is governed by a BSD-style license 4 // that can be found in the COPYING file in the root of the source 5 // tree. An additional intellectual property rights grant can be found 6 // in the file PATENTS. All contributing project authors may 7 // be found in the AUTHORS file in the root of the source tree. 8 // ----------------------------------------------------------------------------- 9 // 10 // SSE2 variant of methods for lossless decoder 11 // 12 // Author: Skal (pascal.massimino (at) gmail.com) 13 14 #include "./dsp.h" 15 16 #if defined(WEBP_USE_SSE2) 17 #include <assert.h> 18 #include <emmintrin.h> 19 #include "./lossless.h" 20 21 //------------------------------------------------------------------------------ 22 // Predictor Transform 23 24 static WEBP_INLINE uint32_t ClampedAddSubtractFull(uint32_t c0, uint32_t c1, 25 uint32_t c2) { 26 const __m128i zero = _mm_setzero_si128(); 27 const __m128i C0 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c0), zero); 28 const __m128i C1 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c1), zero); 29 const __m128i C2 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c2), zero); 30 const __m128i V1 = _mm_add_epi16(C0, C1); 31 const __m128i V2 = _mm_sub_epi16(V1, C2); 32 const __m128i b = _mm_packus_epi16(V2, V2); 33 const uint32_t output = _mm_cvtsi128_si32(b); 34 return output; 35 } 36 37 static WEBP_INLINE uint32_t ClampedAddSubtractHalf(uint32_t c0, uint32_t c1, 38 uint32_t c2) { 39 const __m128i zero = _mm_setzero_si128(); 40 const __m128i C0 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c0), zero); 41 const __m128i C1 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c1), zero); 42 const __m128i B0 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c2), zero); 43 const __m128i avg = _mm_add_epi16(C1, C0); 44 const __m128i A0 = _mm_srli_epi16(avg, 1); 45 const __m128i A1 = _mm_sub_epi16(A0, B0); 46 const __m128i BgtA = _mm_cmpgt_epi16(B0, A0); 47 const __m128i A2 = _mm_sub_epi16(A1, BgtA); 48 const __m128i A3 = _mm_srai_epi16(A2, 1); 49 const __m128i A4 = _mm_add_epi16(A0, A3); 50 const __m128i A5 = _mm_packus_epi16(A4, A4); 51 const uint32_t output = _mm_cvtsi128_si32(A5); 52 return output; 53 } 54 55 static WEBP_INLINE uint32_t Select(uint32_t a, uint32_t b, uint32_t c) { 56 int pa_minus_pb; 57 const __m128i zero = _mm_setzero_si128(); 58 const __m128i A0 = _mm_cvtsi32_si128(a); 59 const __m128i B0 = _mm_cvtsi32_si128(b); 60 const __m128i C0 = _mm_cvtsi32_si128(c); 61 const __m128i AC0 = _mm_subs_epu8(A0, C0); 62 const __m128i CA0 = _mm_subs_epu8(C0, A0); 63 const __m128i BC0 = _mm_subs_epu8(B0, C0); 64 const __m128i CB0 = _mm_subs_epu8(C0, B0); 65 const __m128i AC = _mm_or_si128(AC0, CA0); 66 const __m128i BC = _mm_or_si128(BC0, CB0); 67 const __m128i pa = _mm_unpacklo_epi8(AC, zero); // |a - c| 68 const __m128i pb = _mm_unpacklo_epi8(BC, zero); // |b - c| 69 const __m128i diff = _mm_sub_epi16(pb, pa); 70 { 71 int16_t out[8]; 72 _mm_storeu_si128((__m128i*)out, diff); 73 pa_minus_pb = out[0] + out[1] + out[2] + out[3]; 74 } 75 return (pa_minus_pb <= 0) ? a : b; 76 } 77 78 static WEBP_INLINE __m128i Average2_128i(uint32_t a0, uint32_t a1) { 79 const __m128i zero = _mm_setzero_si128(); 80 const __m128i A0 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(a0), zero); 81 const __m128i A1 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(a1), zero); 82 const __m128i sum = _mm_add_epi16(A1, A0); 83 const __m128i avg = _mm_srli_epi16(sum, 1); 84 return avg; 85 } 86 87 static WEBP_INLINE uint32_t Average2(uint32_t a0, uint32_t a1) { 88 const __m128i avg = Average2_128i(a0, a1); 89 const __m128i A2 = _mm_packus_epi16(avg, avg); 90 const uint32_t output = _mm_cvtsi128_si32(A2); 91 return output; 92 } 93 94 static WEBP_INLINE uint32_t Average3(uint32_t a0, uint32_t a1, uint32_t a2) { 95 const __m128i zero = _mm_setzero_si128(); 96 const __m128i avg1 = Average2_128i(a0, a2); 97 const __m128i A1 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(a1), zero); 98 const __m128i sum = _mm_add_epi16(avg1, A1); 99 const __m128i avg2 = _mm_srli_epi16(sum, 1); 100 const __m128i A2 = _mm_packus_epi16(avg2, avg2); 101 const uint32_t output = _mm_cvtsi128_si32(A2); 102 return output; 103 } 104 105 static WEBP_INLINE uint32_t Average4(uint32_t a0, uint32_t a1, 106 uint32_t a2, uint32_t a3) { 107 const __m128i avg1 = Average2_128i(a0, a1); 108 const __m128i avg2 = Average2_128i(a2, a3); 109 const __m128i sum = _mm_add_epi16(avg2, avg1); 110 const __m128i avg3 = _mm_srli_epi16(sum, 1); 111 const __m128i A0 = _mm_packus_epi16(avg3, avg3); 112 const uint32_t output = _mm_cvtsi128_si32(A0); 113 return output; 114 } 115 116 static uint32_t Predictor5(uint32_t left, const uint32_t* const top) { 117 const uint32_t pred = Average3(left, top[0], top[1]); 118 return pred; 119 } 120 static uint32_t Predictor6(uint32_t left, const uint32_t* const top) { 121 const uint32_t pred = Average2(left, top[-1]); 122 return pred; 123 } 124 static uint32_t Predictor7(uint32_t left, const uint32_t* const top) { 125 const uint32_t pred = Average2(left, top[0]); 126 return pred; 127 } 128 static uint32_t Predictor8(uint32_t left, const uint32_t* const top) { 129 const uint32_t pred = Average2(top[-1], top[0]); 130 (void)left; 131 return pred; 132 } 133 static uint32_t Predictor9(uint32_t left, const uint32_t* const top) { 134 const uint32_t pred = Average2(top[0], top[1]); 135 (void)left; 136 return pred; 137 } 138 static uint32_t Predictor10(uint32_t left, const uint32_t* const top) { 139 const uint32_t pred = Average4(left, top[-1], top[0], top[1]); 140 return pred; 141 } 142 static uint32_t Predictor11(uint32_t left, const uint32_t* const top) { 143 const uint32_t pred = Select(top[0], left, top[-1]); 144 return pred; 145 } 146 static uint32_t Predictor12(uint32_t left, const uint32_t* const top) { 147 const uint32_t pred = ClampedAddSubtractFull(left, top[0], top[-1]); 148 return pred; 149 } 150 static uint32_t Predictor13(uint32_t left, const uint32_t* const top) { 151 const uint32_t pred = ClampedAddSubtractHalf(left, top[0], top[-1]); 152 return pred; 153 } 154 155 //------------------------------------------------------------------------------ 156 // Subtract-Green Transform 157 158 static void AddGreenToBlueAndRed(uint32_t* argb_data, int num_pixels) { 159 int i; 160 for (i = 0; i + 4 <= num_pixels; i += 4) { 161 const __m128i in = _mm_loadu_si128((__m128i*)&argb_data[i]); // argb 162 const __m128i A = _mm_srli_epi16(in, 8); // 0 a 0 g 163 const __m128i B = _mm_shufflelo_epi16(A, _MM_SHUFFLE(2, 2, 0, 0)); 164 const __m128i C = _mm_shufflehi_epi16(B, _MM_SHUFFLE(2, 2, 0, 0)); // 0g0g 165 const __m128i out = _mm_add_epi8(in, C); 166 _mm_storeu_si128((__m128i*)&argb_data[i], out); 167 } 168 // fallthrough and finish off with plain-C 169 VP8LAddGreenToBlueAndRed_C(argb_data + i, num_pixels - i); 170 } 171 172 //------------------------------------------------------------------------------ 173 // Color Transform 174 175 static void TransformColorInverse(const VP8LMultipliers* const m, 176 uint32_t* argb_data, int num_pixels) { 177 // sign-extended multiplying constants, pre-shifted by 5. 178 #define CST(X) (((int16_t)(m->X << 8)) >> 5) // sign-extend 179 const __m128i mults_rb = _mm_set_epi16( 180 CST(green_to_red_), CST(green_to_blue_), 181 CST(green_to_red_), CST(green_to_blue_), 182 CST(green_to_red_), CST(green_to_blue_), 183 CST(green_to_red_), CST(green_to_blue_)); 184 const __m128i mults_b2 = _mm_set_epi16( 185 CST(red_to_blue_), 0, CST(red_to_blue_), 0, 186 CST(red_to_blue_), 0, CST(red_to_blue_), 0); 187 #undef CST 188 const __m128i mask_ag = _mm_set1_epi32(0xff00ff00); // alpha-green masks 189 int i; 190 for (i = 0; i + 4 <= num_pixels; i += 4) { 191 const __m128i in = _mm_loadu_si128((__m128i*)&argb_data[i]); // argb 192 const __m128i A = _mm_and_si128(in, mask_ag); // a 0 g 0 193 const __m128i B = _mm_shufflelo_epi16(A, _MM_SHUFFLE(2, 2, 0, 0)); 194 const __m128i C = _mm_shufflehi_epi16(B, _MM_SHUFFLE(2, 2, 0, 0)); // g0g0 195 const __m128i D = _mm_mulhi_epi16(C, mults_rb); // x dr x db1 196 const __m128i E = _mm_add_epi8(in, D); // x r' x b' 197 const __m128i F = _mm_slli_epi16(E, 8); // r' 0 b' 0 198 const __m128i G = _mm_mulhi_epi16(F, mults_b2); // x db2 0 0 199 const __m128i H = _mm_srli_epi32(G, 8); // 0 x db2 0 200 const __m128i I = _mm_add_epi8(H, F); // r' x b'' 0 201 const __m128i J = _mm_srli_epi16(I, 8); // 0 r' 0 b'' 202 const __m128i out = _mm_or_si128(J, A); 203 _mm_storeu_si128((__m128i*)&argb_data[i], out); 204 } 205 // Fall-back to C-version for left-overs. 206 VP8LTransformColorInverse_C(m, argb_data + i, num_pixels - i); 207 } 208 209 //------------------------------------------------------------------------------ 210 // Color-space conversion functions 211 212 static void ConvertBGRAToRGBA(const uint32_t* src, 213 int num_pixels, uint8_t* dst) { 214 const __m128i* in = (const __m128i*)src; 215 __m128i* out = (__m128i*)dst; 216 while (num_pixels >= 8) { 217 const __m128i bgra0 = _mm_loadu_si128(in++); // bgra0|bgra1|bgra2|bgra3 218 const __m128i bgra4 = _mm_loadu_si128(in++); // bgra4|bgra5|bgra6|bgra7 219 const __m128i v0l = _mm_unpacklo_epi8(bgra0, bgra4); // b0b4g0g4r0r4a0a4... 220 const __m128i v0h = _mm_unpackhi_epi8(bgra0, bgra4); // b2b6g2g6r2r6a2a6... 221 const __m128i v1l = _mm_unpacklo_epi8(v0l, v0h); // b0b2b4b6g0g2g4g6... 222 const __m128i v1h = _mm_unpackhi_epi8(v0l, v0h); // b1b3b5b7g1g3g5g7... 223 const __m128i v2l = _mm_unpacklo_epi8(v1l, v1h); // b0...b7 | g0...g7 224 const __m128i v2h = _mm_unpackhi_epi8(v1l, v1h); // r0...r7 | a0...a7 225 const __m128i ga0 = _mm_unpackhi_epi64(v2l, v2h); // g0...g7 | a0...a7 226 const __m128i rb0 = _mm_unpacklo_epi64(v2h, v2l); // r0...r7 | b0...b7 227 const __m128i rg0 = _mm_unpacklo_epi8(rb0, ga0); // r0g0r1g1 ... r6g6r7g7 228 const __m128i ba0 = _mm_unpackhi_epi8(rb0, ga0); // b0a0b1a1 ... b6a6b7a7 229 const __m128i rgba0 = _mm_unpacklo_epi16(rg0, ba0); // rgba0|rgba1... 230 const __m128i rgba4 = _mm_unpackhi_epi16(rg0, ba0); // rgba4|rgba5... 231 _mm_storeu_si128(out++, rgba0); 232 _mm_storeu_si128(out++, rgba4); 233 num_pixels -= 8; 234 } 235 // left-overs 236 VP8LConvertBGRAToRGBA_C((const uint32_t*)in, num_pixels, (uint8_t*)out); 237 } 238 239 static void ConvertBGRAToRGBA4444(const uint32_t* src, 240 int num_pixels, uint8_t* dst) { 241 const __m128i mask_0x0f = _mm_set1_epi8(0x0f); 242 const __m128i mask_0xf0 = _mm_set1_epi8(0xf0); 243 const __m128i* in = (const __m128i*)src; 244 __m128i* out = (__m128i*)dst; 245 while (num_pixels >= 8) { 246 const __m128i bgra0 = _mm_loadu_si128(in++); // bgra0|bgra1|bgra2|bgra3 247 const __m128i bgra4 = _mm_loadu_si128(in++); // bgra4|bgra5|bgra6|bgra7 248 const __m128i v0l = _mm_unpacklo_epi8(bgra0, bgra4); // b0b4g0g4r0r4a0a4... 249 const __m128i v0h = _mm_unpackhi_epi8(bgra0, bgra4); // b2b6g2g6r2r6a2a6... 250 const __m128i v1l = _mm_unpacklo_epi8(v0l, v0h); // b0b2b4b6g0g2g4g6... 251 const __m128i v1h = _mm_unpackhi_epi8(v0l, v0h); // b1b3b5b7g1g3g5g7... 252 const __m128i v2l = _mm_unpacklo_epi8(v1l, v1h); // b0...b7 | g0...g7 253 const __m128i v2h = _mm_unpackhi_epi8(v1l, v1h); // r0...r7 | a0...a7 254 const __m128i ga0 = _mm_unpackhi_epi64(v2l, v2h); // g0...g7 | a0...a7 255 const __m128i rb0 = _mm_unpacklo_epi64(v2h, v2l); // r0...r7 | b0...b7 256 const __m128i ga1 = _mm_srli_epi16(ga0, 4); // g0-|g1-|...|a6-|a7- 257 const __m128i rb1 = _mm_and_si128(rb0, mask_0xf0); // -r0|-r1|...|-b6|-a7 258 const __m128i ga2 = _mm_and_si128(ga1, mask_0x0f); // g0-|g1-|...|a6-|a7- 259 const __m128i rgba0 = _mm_or_si128(ga2, rb1); // rg0..rg7 | ba0..ba7 260 const __m128i rgba1 = _mm_srli_si128(rgba0, 8); // ba0..ba7 | 0 261 #ifdef WEBP_SWAP_16BIT_CSP 262 const __m128i rgba = _mm_unpacklo_epi8(rgba1, rgba0); // barg0...barg7 263 #else 264 const __m128i rgba = _mm_unpacklo_epi8(rgba0, rgba1); // rgba0...rgba7 265 #endif 266 _mm_storeu_si128(out++, rgba); 267 num_pixels -= 8; 268 } 269 // left-overs 270 VP8LConvertBGRAToRGBA4444_C((const uint32_t*)in, num_pixels, (uint8_t*)out); 271 } 272 273 static void ConvertBGRAToRGB565(const uint32_t* src, 274 int num_pixels, uint8_t* dst) { 275 const __m128i mask_0xe0 = _mm_set1_epi8(0xe0); 276 const __m128i mask_0xf8 = _mm_set1_epi8(0xf8); 277 const __m128i mask_0x07 = _mm_set1_epi8(0x07); 278 const __m128i* in = (const __m128i*)src; 279 __m128i* out = (__m128i*)dst; 280 while (num_pixels >= 8) { 281 const __m128i bgra0 = _mm_loadu_si128(in++); // bgra0|bgra1|bgra2|bgra3 282 const __m128i bgra4 = _mm_loadu_si128(in++); // bgra4|bgra5|bgra6|bgra7 283 const __m128i v0l = _mm_unpacklo_epi8(bgra0, bgra4); // b0b4g0g4r0r4a0a4... 284 const __m128i v0h = _mm_unpackhi_epi8(bgra0, bgra4); // b2b6g2g6r2r6a2a6... 285 const __m128i v1l = _mm_unpacklo_epi8(v0l, v0h); // b0b2b4b6g0g2g4g6... 286 const __m128i v1h = _mm_unpackhi_epi8(v0l, v0h); // b1b3b5b7g1g3g5g7... 287 const __m128i v2l = _mm_unpacklo_epi8(v1l, v1h); // b0...b7 | g0...g7 288 const __m128i v2h = _mm_unpackhi_epi8(v1l, v1h); // r0...r7 | a0...a7 289 const __m128i ga0 = _mm_unpackhi_epi64(v2l, v2h); // g0...g7 | a0...a7 290 const __m128i rb0 = _mm_unpacklo_epi64(v2h, v2l); // r0...r7 | b0...b7 291 const __m128i rb1 = _mm_and_si128(rb0, mask_0xf8); // -r0..-r7|-b0..-b7 292 const __m128i g_lo1 = _mm_srli_epi16(ga0, 5); 293 const __m128i g_lo2 = _mm_and_si128(g_lo1, mask_0x07); // g0-...g7-|xx (3b) 294 const __m128i g_hi1 = _mm_slli_epi16(ga0, 3); 295 const __m128i g_hi2 = _mm_and_si128(g_hi1, mask_0xe0); // -g0...-g7|xx (3b) 296 const __m128i b0 = _mm_srli_si128(rb1, 8); // -b0...-b7|0 297 const __m128i rg1 = _mm_or_si128(rb1, g_lo2); // gr0...gr7|xx 298 const __m128i b1 = _mm_srli_epi16(b0, 3); 299 const __m128i gb1 = _mm_or_si128(b1, g_hi2); // bg0...bg7|xx 300 #ifdef WEBP_SWAP_16BIT_CSP 301 const __m128i rgba = _mm_unpacklo_epi8(gb1, rg1); // rggb0...rggb7 302 #else 303 const __m128i rgba = _mm_unpacklo_epi8(rg1, gb1); // bgrb0...bgrb7 304 #endif 305 _mm_storeu_si128(out++, rgba); 306 num_pixels -= 8; 307 } 308 // left-overs 309 VP8LConvertBGRAToRGB565_C((const uint32_t*)in, num_pixels, (uint8_t*)out); 310 } 311 312 static void ConvertBGRAToBGR(const uint32_t* src, 313 int num_pixels, uint8_t* dst) { 314 const __m128i mask_l = _mm_set_epi32(0, 0x00ffffff, 0, 0x00ffffff); 315 const __m128i mask_h = _mm_set_epi32(0x00ffffff, 0, 0x00ffffff, 0); 316 const __m128i* in = (const __m128i*)src; 317 const uint8_t* const end = dst + num_pixels * 3; 318 // the last storel_epi64 below writes 8 bytes starting at offset 18 319 while (dst + 26 <= end) { 320 const __m128i bgra0 = _mm_loadu_si128(in++); // bgra0|bgra1|bgra2|bgra3 321 const __m128i bgra4 = _mm_loadu_si128(in++); // bgra4|bgra5|bgra6|bgra7 322 const __m128i a0l = _mm_and_si128(bgra0, mask_l); // bgr0|0|bgr0|0 323 const __m128i a4l = _mm_and_si128(bgra4, mask_l); // bgr0|0|bgr0|0 324 const __m128i a0h = _mm_and_si128(bgra0, mask_h); // 0|bgr0|0|bgr0 325 const __m128i a4h = _mm_and_si128(bgra4, mask_h); // 0|bgr0|0|bgr0 326 const __m128i b0h = _mm_srli_epi64(a0h, 8); // 000b|gr00|000b|gr00 327 const __m128i b4h = _mm_srli_epi64(a4h, 8); // 000b|gr00|000b|gr00 328 const __m128i c0 = _mm_or_si128(a0l, b0h); // rgbrgb00|rgbrgb00 329 const __m128i c4 = _mm_or_si128(a4l, b4h); // rgbrgb00|rgbrgb00 330 const __m128i c2 = _mm_srli_si128(c0, 8); 331 const __m128i c6 = _mm_srli_si128(c4, 8); 332 _mm_storel_epi64((__m128i*)(dst + 0), c0); 333 _mm_storel_epi64((__m128i*)(dst + 6), c2); 334 _mm_storel_epi64((__m128i*)(dst + 12), c4); 335 _mm_storel_epi64((__m128i*)(dst + 18), c6); 336 dst += 24; 337 num_pixels -= 8; 338 } 339 // left-overs 340 VP8LConvertBGRAToBGR_C((const uint32_t*)in, num_pixels, dst); 341 } 342 343 //------------------------------------------------------------------------------ 344 // Entry point 345 346 extern void VP8LDspInitSSE2(void); 347 348 WEBP_TSAN_IGNORE_FUNCTION void VP8LDspInitSSE2(void) { 349 VP8LPredictors[5] = Predictor5; 350 VP8LPredictors[6] = Predictor6; 351 VP8LPredictors[7] = Predictor7; 352 VP8LPredictors[8] = Predictor8; 353 VP8LPredictors[9] = Predictor9; 354 VP8LPredictors[10] = Predictor10; 355 VP8LPredictors[11] = Predictor11; 356 VP8LPredictors[12] = Predictor12; 357 VP8LPredictors[13] = Predictor13; 358 359 VP8LAddGreenToBlueAndRed = AddGreenToBlueAndRed; 360 VP8LTransformColorInverse = TransformColorInverse; 361 362 VP8LConvertBGRAToRGBA = ConvertBGRAToRGBA; 363 VP8LConvertBGRAToRGBA4444 = ConvertBGRAToRGBA4444; 364 VP8LConvertBGRAToRGB565 = ConvertBGRAToRGB565; 365 VP8LConvertBGRAToBGR = ConvertBGRAToBGR; 366 } 367 368 #else // !WEBP_USE_SSE2 369 370 WEBP_DSP_INIT_STUB(VP8LDspInitSSE2) 371 372 #endif // WEBP_USE_SSE2 373